Refine
Year of publication
Document Type
- Article (36)
- Postprint (16)
- Other (2)
- Part of a Book (1)
- Habilitation Thesis (1)
Is part of the Bibliography
- yes (56)
Keywords
- Algorithm (2)
- Band (2)
- Kwajalein (2)
- Methodology (2)
- Nordic catchments (2)
- Reflectivity (2)
- TELEMAC-2D model (2)
- Uncertainties (2)
- Urban pluvial flood susceptibility (2)
- Weather (2)
Transferability of data-driven models to predict urban pluvial flood water depth in Berlin, Germany
(2023)
Data-driven models have been recently suggested to surrogate computationally expensive hydrodynamic models to map flood hazards. However, most studies focused on developing models for the same area or the same precipitation event. It is thus not obvious how transferable the models are in space. This study evaluates the performance of a convolutional neural network (CNN) based on the U-Net architecture and the random forest (RF) algorithm to predict flood water depth, the models' transferability in space and performance improvement using transfer learning techniques. We used three study areas in Berlin to train, validate and test the models. The results showed that (1) the RF models outperformed the CNN models for predictions within the training domain, presumable at the cost of overfitting; (2) the CNN models had significantly higher potential than the RF models to generalize beyond the training domain; and (3) the CNN models could better benefit from transfer learning technique to boost their performance outside training domains than RF models.
Transferability of data-driven models to predict urban pluvial flood water depth in Berlin, Germany
(2023)
Data-driven models have been recently suggested to surrogate computationally expensive hydrodynamic models to map flood hazards. However, most studies focused on developing models for the same area or the same precipitation event. It is thus not obvious how transferable the models are in space. This study evaluates the performance of a convolutional neural network (CNN) based on the U-Net architecture and the random forest (RF) algorithm to predict flood water depth, the models' transferability in space and performance improvement using transfer learning techniques. We used three study areas in Berlin to train, validate and test the models. The results showed that (1) the RF models outperformed the CNN models for predictions within the training domain, presumable at the cost of overfitting; (2) the CNN models had significantly higher potential than the RF models to generalize beyond the training domain; and (3) the CNN models could better benefit from transfer learning technique to boost their performance outside training domains than RF models.
Quantifying the extremeness of heavy precipitation allows for the comparison of events. Conventional quantitative indices, however, typically neglect the spatial extent or the duration, while both are important to understand potential impacts. In 2014, the weather extremity index (WEI) was suggested to quantify the extremeness of an event and to identify the spatial and temporal scale at which the event was most extreme. However, the WEI does not account for the fact that one event can be extreme at various spatial and temporal scales. To better understand and detect the compound nature of precipitation events, we suggest complementing the original WEI with a “cross-scale weather extremity index” (xWEI), which integrates extremeness over relevant scales instead of determining its maximum.
Based on a set of 101 extreme precipitation events in Germany, we outline and demonstrate the computation of both WEI and xWEI. We find that the choice of the index can lead to considerable differences in the assessment of past events but that the most extreme events are ranked consistently, independently of the index. Even then, the xWEI can reveal cross-scale properties which would otherwise remain hidden. This also applies to the disastrous event from July 2021, which clearly outranks all other analyzed events with regard to both WEI and xWEI.
While demonstrating the added value of xWEI, we also identify various methodological challenges along the required computational workflow: these include the parameter estimation for the extreme value distributions, the definition of maximum spatial extent and temporal duration, and the weighting of extremeness at different scales. These challenges, however, also represent opportunities to adjust the retrieval of WEI and xWEI to specific user requirements and application scenarios.
Quantifying the extremeness of heavy precipitation allows for the comparison of events. Conventional quantitative indices, however, typically neglect the spatial extent or the duration, while both are important to understand potential impacts. In 2014, the weather extremity index (WEI) was suggested to quantify the extremeness of an event and to identify the spatial and temporal scale at which the event was most extreme. However, the WEI does not account for the fact that one event can be extreme at various spatial and temporal scales. To better understand and detect the compound nature of precipitation events, we suggest complementing the original WEI with a “cross-scale weather extremity index” (xWEI), which integrates extremeness over relevant scales instead of determining its maximum.
Based on a set of 101 extreme precipitation events in Germany, we outline and demonstrate the computation of both WEI and xWEI. We find that the choice of the index can lead to considerable differences in the assessment of past events but that the most extreme events are ranked consistently, independently of the index. Even then, the xWEI can reveal cross-scale properties which would otherwise remain hidden. This also applies to the disastrous event from July 2021, which clearly outranks all other analyzed events with regard to both WEI and xWEI.
While demonstrating the added value of xWEI, we also identify various methodological challenges along the required computational workflow: these include the parameter estimation for the extreme value distributions, the definition of maximum spatial extent and temporal duration, and the weighting of extremeness at different scales. These challenges, however, also represent opportunities to adjust the retrieval of WEI and xWEI to specific user requirements and application scenarios.
Identifying urban pluvial flood-prone areas is necessary but the application of two-dimensional hydrodynamic models is limited to small areas. Data-driven models have been showing their ability to map flood susceptibility but their application in urban pluvial flooding is still rare. A flood inventory (4333 flooded locations) and 11 factors which potentially indicate an increased hazard for pluvial flooding were used to implement convolutional neural network (CNN), artificial neural network (ANN), random forest (RF) and support vector machine (SVM) to: (1) Map flood susceptibility in Berlin at 30, 10, 5, and 2 m spatial resolutions. (2) Evaluate the trained models' transferability in space. (3) Estimate the most useful factors for flood susceptibility mapping. The models' performance was validated using the Kappa, and the area under the receiver operating characteristic curve (AUC). The results indicated that all models perform very well (minimum AUC = 0.87 for the testing dataset). The RF models outperformed all other models at all spatial resolutions and the RF model at 2 m spatial resolution was superior for the present flood inventory and predictor variables. The majority of the models had a moderate performance for predictions outside the training area based on Kappa evaluation (minimum AUC = 0.8). Aspect and altitude were the most influencing factors on the image-based and point-based models respectively. Data-driven models can be a reliable tool for urban pluvial flood susceptibility mapping wherever a reliable flood inventory is available.
Identifying urban pluvial flood-prone areas is necessary but the application of two-dimensional hydrodynamic models is limited to small areas. Data-driven models have been showing their ability to map flood susceptibility but their application in urban pluvial flooding is still rare. A flood inventory (4333 flooded locations) and 11 factors which potentially indicate an increased hazard for pluvial flooding were used to implement convolutional neural network (CNN), artificial neural network (ANN), random forest (RF) and support vector machine (SVM) to: (1) Map flood susceptibility in Berlin at 30, 10, 5, and 2 m spatial resolutions. (2) Evaluate the trained models' transferability in space. (3) Estimate the most useful factors for flood susceptibility mapping. The models' performance was validated using the Kappa, and the area under the receiver operating characteristic curve (AUC). The results indicated that all models perform very well (minimum AUC = 0.87 for the testing dataset). The RF models outperformed all other models at all spatial resolutions and the RF model at 2 m spatial resolution was superior for the present flood inventory and predictor variables. The majority of the models had a moderate performance for predictions outside the training area based on Kappa evaluation (minimum AUC = 0.8). Aspect and altitude were the most influencing factors on the image-based and point-based models respectively. Data-driven models can be a reliable tool for urban pluvial flood susceptibility mapping wherever a reliable flood inventory is available.
Cosmic-ray neutron sensing (CRNS) is a non-invasive tool for measuring hydrogen pools such as soil moisture, snow or vegetation. The intrinsic integration over a radial hectare-scale footprint is a clear advantage for averaging out small-scale heterogeneity, but on the other hand the data may become hard to interpret in complex terrain with patchy land use.
This study presents a directional shielding approach to prevent neutrons from certain angles from being counted while counting neutrons entering the detector from other angles and explores its potential to gain a sharper horizontal view on the surrounding soil moisture distribution.
Using the Monte Carlo code URANOS (Ultra Rapid Neutron-Only Simulation), we modelled the effect of additional polyethylene shields on the horizontal field of view and assessed its impact on the epithermal count rate, propagated uncertainties and aggregation time.
The results demonstrate that directional CRNS measurements are strongly dominated by isotropic neutron transport, which dilutes the signal of the targeted direction especially from the far field. For typical count rates of customary CRNS stations, directional shielding of half-spaces could not lead to acceptable precision at a daily time resolution. However, the mere statistical distinction of two rates should be feasible.
Cosmic-ray neutron sensing (CRNS) has become an effective method to measure soil moisture at a horizontal scale of hundreds of metres and a depth of decimetres. Recent studies proposed operating CRNS in a network with overlapping footprints in order to cover root-zone water dynamics at the small catchment scale and, at the same time, to represent spatial heterogeneity. In a joint field campaign from September to November 2020 (JFC-2020), five German research institutions deployed 15 CRNS sensors in the 0.4 km2 Wüstebach catchment (Eifel mountains, Germany). The catchment is dominantly forested (but includes a substantial fraction of open vegetation) and features a topographically distinct catchment boundary. In addition to the dense CRNS coverage, the campaign featured a unique combination of additional instruments and techniques: hydro-gravimetry (to detect water storage dynamics also below the root zone); ground-based and, for the first time, airborne CRNS roving; an extensive wireless soil sensor network, supplemented by manual measurements; and six weighable lysimeters. Together with comprehensive data from the long-term local research infrastructure, the published data set (available at https://doi.org/10.23728/b2share.756ca0485800474e9dc7f5949c63b872; Heistermann et al., 2022) will be a valuable asset in various research contexts: to advance the retrieval of landscape water storage from CRNS, wireless soil sensor networks, or hydrogravimetry; to identify scale-specific combinations of sensors and methods to represent soil moisture variability; to improve the understanding and simulation of land–atmosphere exchange as well as hydrological and hydrogeological processes at the hillslope and the catchment scale; and to support the retrieval of soil water content from airborne and spaceborne remote sensing platforms.
Cosmic-ray neutron sensing (CRNS) is a non-invasive tool for measuring hydrogen pools such as soil moisture, snow or vegetation. The intrinsic integration over a radial hectare-scale footprint is a clear advantage for averaging out small-scale heterogeneity, but on the other hand the data may become hard to interpret in complex terrain with patchy land use.
This study presents a directional shielding approach to prevent neutrons from certain angles from being counted while counting neutrons entering the detector from other angles and explores its potential to gain a sharper horizontal view on the surrounding soil moisture distribution.
Using the Monte Carlo code URANOS (Ultra Rapid Neutron-Only Simulation), we modelled the effect of additional polyethylene shields on the horizontal field of view and assessed its impact on the epithermal count rate, propagated uncertainties and aggregation time.
The results demonstrate that directional CRNS measurements are strongly dominated by isotropic neutron transport, which dilutes the signal of the targeted direction especially from the far field. For typical count rates of customary CRNS stations, directional shielding of half-spaces could not lead to acceptable precision at a daily time resolution. However, the mere statistical distinction of two rates should be feasible.
Cosmic-ray neutron sensing (CRNS) has become an effective method to measure soil moisture at a horizontal scale of hundreds of metres and a depth of decimetres. Recent studies proposed operating CRNS in a network with overlapping footprints in order to cover root-zone water dynamics at the small catchment scale and, at the same time, to represent spatial heterogeneity. In a joint field campaign from September to November 2020 (JFC-2020), five German research institutions deployed 15 CRNS sensors in the 0.4 km2 Wüstebach catchment (Eifel mountains, Germany). The catchment is dominantly forested (but includes a substantial fraction of open vegetation) and features a topographically distinct catchment boundary. In addition to the dense CRNS coverage, the campaign featured a unique combination of additional instruments and techniques: hydro-gravimetry (to detect water storage dynamics also below the root zone); ground-based and, for the first time, airborne CRNS roving; an extensive wireless soil sensor network, supplemented by manual measurements; and six weighable lysimeters. Together with comprehensive data from the long-term local research infrastructure, the published data set (available at https://doi.org/10.23728/b2share.756ca0485800474e9dc7f5949c63b872; Heistermann et al., 2022) will be a valuable asset in various research contexts: to advance the retrieval of landscape water storage from CRNS, wireless soil sensor networks, or hydrogravimetry; to identify scale-specific combinations of sensors and methods to represent soil moisture variability; to improve the understanding and simulation of land–atmosphere exchange as well as hydrological and hydrogeological processes at the hillslope and the catchment scale; and to support the retrieval of soil water content from airborne and spaceborne remote sensing platforms.
We systematically explore the effect of calibration data length on the performance of a conceptual hydrological model, GR4H, in comparison to two Artificial Neural Network (ANN) architectures: Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRU), which have just recently been introduced to the field of hydrology. We implemented a case study for six river basins across the contiguous United States, with 25 years of meteorological and discharge data. Nine years were reserved for independent validation; two years were used as a warm-up period, one year for each of the calibration and validation periods, respectively; from the remaining 14 years, we sampled increasing amounts of data for model calibration, and found pronounced differences in model performance. While GR4H required less data to converge, LSTM and GRU caught up at a remarkable rate, considering their number of parameters. Also, LSTM and GRU exhibited the higher calibration instability in comparison to GR4H. These findings confirm the potential of modern deep-learning architectures in rainfall runoff modelling, but also highlight the noticeable differences between them in regard to the effect of calibration data length.
The presence of impermeable surfaces in urban areas hinders natural drainage and directs the surface runoff to storm drainage systems with finite capacity, which makes these areas prone to pluvial flooding. The occurrence of pluvial flooding depends on the existence of minimal areas for surface runoff generation and concentration. Detailed hydrologic and hydrodynamic simulations are computationally expensive and require intensive resources. This study compared and evaluated the performance of two simplified methods to identify urban pluvial flood-prone areas, namely the fill–spill–merge (FSM) method and the topographic wetness index (TWI) method and used the TELEMAC-2D hydrodynamic numerical model for benchmarking and validation. The FSM method uses common GIS operations to identify flood-prone depressions from a high-resolution digital elevation model (DEM). The TWI method employs the maximum likelihood method (MLE) to probabilistically calibrate a TWI threshold (τ) based on the inundation maps from a 2D hydrodynamic model for a given spatial window (W) within the urban area. We found that the FSM method clearly outperforms the TWI method both conceptually and effectively in terms of model performance.
The presence of impermeable surfaces in urban areas hinders natural drainage and directs the surface runoff to storm drainage systems with finite capacity, which makes these areas prone to pluvial flooding. The occurrence of pluvial flooding depends on the existence of minimal areas for surface runoff generation and concentration. Detailed hydrologic and hydrodynamic simulations are computationally expensive and require intensive resources. This study compared and evaluated the performance of two simplified methods to identify urban pluvial flood-prone areas, namely the fill–spill–merge (FSM) method and the topographic wetness index (TWI) method and used the TELEMAC-2D hydrodynamic numerical model for benchmarking and validation. The FSM method uses common GIS operations to identify flood-prone depressions from a high-resolution digital elevation model (DEM). The TWI method employs the maximum likelihood method (MLE) to probabilistically calibrate a TWI threshold (τ) based on the inundation maps from a 2D hydrodynamic model for a given spatial window (W) within the urban area. We found that the FSM method clearly outperforms the TWI method both conceptually and effectively in terms of model performance.
Cosmic-ray neutron sensing (CRNS) is a powerful technique for retrieving representative estimates of soil water content at a horizontal scale of hectometres (the “field scale”) and depths of tens of centimetres (“the root zone”). This study demonstrates the potential of the CRNS technique to obtain spatio-temporal patterns of soil moisture beyond the integrated volume from isolated CRNS footprints. We use data from an observational campaign carried out between May and July 2019 that featured a dense network of more than 20 neutron detectors with partly overlapping footprints in an area that exhibits pronounced soil moisture gradients within one square kilometre. The present study is the first to combine these observations in order to represent the heterogeneity of soil water content at the sub-footprint scale as well as between the CRNS stations. First, we apply a state-of-the-art procedure to correct the observed neutron count rates for static effects (heterogeneity in space, e.g. soil organic matter) and dynamic effects (heterogeneity in time, e.g. barometric pressure). Based on the homogenized neutron data, we investigate the robustness of a calibration approach that uses a single calibration parameter across all CRNS stations. Finally, we benchmark two different interpolation techniques for obtaining spatio-temporal representations of soil moisture: first, ordinary Kriging with a fixed range; second, spatial interpolation complemented by geophysical inversion (“constrained interpolation”). To that end, we optimize the parameters of a geostatistical interpolation model so that the error in the forward-simulated neutron count rates is minimized, and suggest a heuristic forward operator to make the optimization problem computationally feasible. Comparison with independent measurements from a cluster of soil moisture sensors (SoilNet) shows that the constrained interpolation approach is superior for representing horizontal soil moisture gradients at the hectometre scale. The study demonstrates how a CRNS network can be used to generate coherent, consistent, and continuous soil moisture patterns that could be used to validate hydrological models or remote sensing products.
Cosmic-ray neutron sensing (CRNS) is a powerful technique for retrieving representative estimates of soil water content at a horizontal scale of hectometres (the “field scale”) and depths of tens of centimetres (“the root zone”). This study demonstrates the potential of the CRNS technique to obtain spatio-temporal patterns of soil moisture beyond the integrated volume from isolated CRNS footprints. We use data from an observational campaign carried out between May and July 2019 that featured a dense network of more than 20 neutron detectors with partly overlapping footprints in an area that exhibits pronounced soil moisture gradients within one square kilometre. The present study is the first to combine these observations in order to represent the heterogeneity of soil water content at the sub-footprint scale as well as between the CRNS stations. First, we apply a state-of-the-art procedure to correct the observed neutron count rates for static effects (heterogeneity in space, e.g. soil organic matter) and dynamic effects (heterogeneity in time, e.g. barometric pressure). Based on the homogenized neutron data, we investigate the robustness of a calibration approach that uses a single calibration parameter across all CRNS stations. Finally, we benchmark two different interpolation techniques for obtaining spatio-temporal representations of soil moisture: first, ordinary Kriging with a fixed range; second, spatial interpolation complemented by geophysical inversion (“constrained interpolation”). To that end, we optimize the parameters of a geostatistical interpolation model so that the error in the forward-simulated neutron count rates is minimized, and suggest a heuristic forward operator to make the optimization problem computationally feasible. Comparison with independent measurements from a cluster of soil moisture sensors (SoilNet) shows that the constrained interpolation approach is superior for representing horizontal soil moisture gradients at the hectometre scale. The study demonstrates how a CRNS network can be used to generate coherent, consistent, and continuous soil moisture patterns that could be used to validate hydrological models or remote sensing products.
In recent years, urban and rural flash floods in Europe and abroad have gained considerable attention because of their sudden occurrence, severe material damages and even danger to life of inhabitants. This contribution addresses questions about possibly changing environmental conditions which might have altered the occurrence frequencies of such events and their consequences. We analyze the following major fields of environmental changes.
Altered high intensity rain storm conditions, as a consequence of regionalwarming; Possibly altered runoff generation conditions in response to high intensity rainfall events; Possibly altered runoff concentration conditions in response to the usage and management of the landscape, such as agricultural, forest practices or rural roads; Effects of engineering measures in the catchment, such as retention basins, check dams, culverts, or river and geomorphological engineering measures.
We take the flash-flood in Braunsbach, SW-Germany, as an example, where a particularly concise flash flood event occurred at the end of May 2016. This extreme cascading natural event led to immense damage in this particular village. The event is retrospectively analyzed with regard to meteorology, hydrology, geomorphology and damage to obtain a quantitative assessment of the processes and their development.
The results show that it was a very rare rainfall event with extreme intensities, which in combination with catchment properties and altered environmental conditions led to extreme runoff, extreme debris flow and immense damages. Due to the complex and interacting processes, no single flood cause can be identified, since only the interplay of those led to such an event. We have shown that environmental changes are important, but-at least for this case study-even natural weather and hydrologic conditions would still have resulted in an extreme flash flood event.
In precipitation nowcasting, it is common to track the motion of precipitation in a sequence of weather radar images and to extrapolate this motion into the future. The total error of such a prediction consists of an error in the predicted location of a precipitation feature and an error in the change of precipitation intensity over lead time. So far, verification measures did not allow isolating the extent of location errors, making it difficult to specifically improve nowcast models with regard to location prediction. In this paper, we introduce a framework to directly quantify the location error. To that end, we detect and track scale-invariant precipitation features (corners) in radar images. We then consider these observed tracks as the true reference in order to evaluate the performance (or, inversely, the error) of any model that aims to predict the future location of a precipitation feature. Hence, the location error of a forecast at any lead time Delta t ahead of the forecast time t corresponds to the Euclidean distance between the observed and the predicted feature locations at t + Delta t. Based on this framework, we carried out a benchmarking case study using one year worth of weather radar composites of the German Weather Service. We evaluated the performance of four extrapolation models, two of which are based on the linear extrapolation of corner motion from t - 1 to t (LK-Lin1) and t - 4 to t (LK-Lin4) and the other two are based on the Dense Inverse Search (DIS) method: motion vectors obtained from DIS are used to predict feature locations by linear (DIS-Lin1) and Semi-Lagrangian extrapolation (DIS-Rot1). Of those four models, DIS-Lin1 and LK-Lin4 turned out to be the most skillful with regard to the prediction of feature location, while we also found that the model skill dramatically depends on the sinuosity of the observed tracks. The dataset of 376,125 detected feature tracks in 2016 is openly available to foster the improvement of location prediction in extrapolation-based nowcasting models.
Many institutions struggle to tap into the potential of their large archives of radar reflectivity: these data are often affected by miscalibration, yet the bias is typically unknown and temporally volatile. Still, relative calibration techniques can be used to correct the measurements a posteriori. For that purpose, the usage of spaceborne reflectivity observations from the Tropical Rainfall Measuring Mission (TRMM) and Global Precipitation Measurement (GPM) platforms has become increasingly popular: the calibration bias of a ground radar (GR) is estimated from its average reflectivity difference to the spaceborne radar (SR). Recently, Crisologo et al. (2018) introduced a formal procedure to enhance the reliability of such estimates: each match between SR and GR observations is assigned a quality index, and the calibration bias is inferred as a quality-weighted average of the differences between SR and GR. The relevance of quality was exemplified for the Subic S-band radar in the Philippines, which is greatly affected by partial beam blockage.
The present study extends the concept of quality-weighted averaging by accounting for path-integrated attenuation (PIA) in addition to beam blockage. This extension becomes vital for radars that operate at the C or X band. Correspondingly, the study setup includes a C-band radar that substantially overlaps with the S-band radar. Based on the extended quality-weighting approach, we retrieve, for each of the two ground radars, a time series of calibration bias estimates from suitable SR overpasses. As a result of applying these estimates to correct the ground radar observations, the consistency between the ground radars in the region of overlap increased substantially. Furthermore, we investigated if the bias estimates can be interpolated in time, so that ground radar observations can be corrected even in the absence of prompt SR overpasses. We found that a moving average approach was most suitable for that purpose, although limited by the absence of explicit records of radar maintenance operations.
RainNet v1.0
(2020)
In this study, we present RainNet, a deep convolutional neural network for radar-based precipitation nowcasting. Its design was inspired by the U-Net and SegNet families of deep learning models, which were originally designed for binary segmentation tasks. RainNet was trained to predict continuous precipitation intensities at a lead time of 5min, using several years of quality-controlled weather radar composites provided by the German Weather Service (DWD). That data set covers Germany with a spatial domain of 900km × 900km and has a resolution of 1km in space and 5min in time. Independent verification experiments were carried out on 11 summer precipitation events from 2016 to 2017. In order to achieve a lead time of 1h, a recursive approach was implemented by using RainNet predictions at 5min lead times as model inputs for longer lead times. In the verification experiments, trivial Eulerian persistence and a conventional model based on optical flow served as benchmarks. The latter is available in the rainymotion library and had previously been shown to outperform DWD's operational nowcasting model for the same set of verification events.
RainNet significantly outperforms the benchmark models at all lead times up to 60min for the routine verification metrics mean absolute error (MAE) and the critical success index (CSI) at intensity thresholds of 0.125, 1, and 5mm h⁻¹. However, rainymotion turned out to be superior in predicting the exceedance of higher intensity thresholds (here 10 and 15mm h⁻¹). The limited ability of RainNet to predict heavy rainfall intensities is an undesirable property which we attribute to a high level of spatial smoothing introduced by the model. At a lead time of 5min, an analysis of power spectral density confirmed a significant loss of spectral power at length scales of 16km and below. Obviously, RainNet had learned an optimal level of smoothing to produce a nowcast at 5min lead time. In that sense, the loss of spectral power at small scales is informative, too, as it reflects the limits of predictability as a function of spatial scale. Beyond the lead time of 5min, however, the increasing level of smoothing is a mere artifact – an analogue to numerical diffusion – that is not a property of RainNet itself but of its recursive application. In the context of early warning, the smoothing is particularly unfavorable since pronounced features of intense precipitation tend to get lost over longer lead times. Hence, we propose several options to address this issue in prospective research, including an adjustment of the loss function for model training, model training for longer lead times, and the prediction of threshold exceedance in terms of a binary segmentation task. Furthermore, we suggest additional input data that could help to better identify situations with imminent precipitation dynamics. The model code, pretrained weights, and training data are provided in open repositories as an input for such future studies.
RainNet v1.0
(2020)
In this study, we present RainNet, a deep convolutional neural network for radar-based precipitation nowcasting. Its design was inspired by the U-Net and SegNet families of deep learning models, which were originally designed for binary segmentation tasks. RainNet was trained to predict continuous precipitation intensities at a lead time of 5min, using several years of quality-controlled weather radar composites provided by the German Weather Service (DWD). That data set covers Germany with a spatial domain of 900km × 900km and has a resolution of 1km in space and 5min in time. Independent verification experiments were carried out on 11 summer precipitation events from 2016 to 2017. In order to achieve a lead time of 1h, a recursive approach was implemented by using RainNet predictions at 5min lead times as model inputs for longer lead times. In the verification experiments, trivial Eulerian persistence and a conventional model based on optical flow served as benchmarks. The latter is available in the rainymotion library and had previously been shown to outperform DWD's operational nowcasting model for the same set of verification events.
RainNet significantly outperforms the benchmark models at all lead times up to 60min for the routine verification metrics mean absolute error (MAE) and the critical success index (CSI) at intensity thresholds of 0.125, 1, and 5mm h⁻¹. However, rainymotion turned out to be superior in predicting the exceedance of higher intensity thresholds (here 10 and 15mm h⁻¹). The limited ability of RainNet to predict heavy rainfall intensities is an undesirable property which we attribute to a high level of spatial smoothing introduced by the model. At a lead time of 5min, an analysis of power spectral density confirmed a significant loss of spectral power at length scales of 16km and below. Obviously, RainNet had learned an optimal level of smoothing to produce a nowcast at 5min lead time. In that sense, the loss of spectral power at small scales is informative, too, as it reflects the limits of predictability as a function of spatial scale. Beyond the lead time of 5min, however, the increasing level of smoothing is a mere artifact – an analogue to numerical diffusion – that is not a property of RainNet itself but of its recursive application. In the context of early warning, the smoothing is particularly unfavorable since pronounced features of intense precipitation tend to get lost over longer lead times. Hence, we propose several options to address this issue in prospective research, including an adjustment of the loss function for model training, model training for longer lead times, and the prediction of threshold exceedance in terms of a binary segmentation task. Furthermore, we suggest additional input data that could help to better identify situations with imminent precipitation dynamics. The model code, pretrained weights, and training data are provided in open repositories as an input for such future studies.